lexigram-ai-rag

Retrieval-Augmented Generation (RAG) pipeline for the Lexigram Framework

Overview

RAG (Retrieval-Augmented Generation) pipeline for the Lexigram Framework. Provides a multi-stage, fully configurable pipeline covering ingestion, query processing, retrieval, context optimisation, synthesis, quality assurance, and post-processing — all wired through the DI container via RAGModule. Zero-config usage starts with sensible defaults.

Full documentation: docs.lexigram.dev

Install

uv add lexigram-ai-rag
# Optional extras
uv add "lexigram-ai-rag[pdf,web]"

Quick Start

from lexigram import Application
from lexigram.di.module import Module, module

from lexigram.ai.rag import RAGModule
from lexigram.ai.rag.config import RAGConfig

@module(imports=[
    RAGModule.configure(
        RAGConfig(
            vector_store_type="pgvector",
            collection_name="my_docs",
            top_k=5,
            enable_citations=True,
        )
    )
])
class AppModule(Module):
    pass

async with Application.boot(modules=[AppModule]) as app:
    # use app.container to resolve services
    ...

Configuration

Zero-config usage: Call RAGModule.configure() with no arguments to use defaults.

Option 1 — YAML file

# application.yaml
ai_rag:
  vector_store_type: "pgvector"
  collection_name: "my_docs"
  top_k: 5
  embedding_model: "text-embedding-ada-002"
  enable_citations: true

Option 2 — Profiles + Environment Variables (recommended)

export LEX_AI_RAG__VECTOR_STORE_TYPE=pgvector
# Environment variables for each field

Option 3 — Python

from lexigram.ai.rag.config import RAGConfig
from lexigram.ai.rag import RAGModule

config = RAGConfig(
    vector_store_type="pgvector",
    collection_name="my_docs",
    top_k=5,
)
RAGModule.configure(config)

Config reference

Module Factory Methods

Key Features

• Multi-stage pipeline: Ingestion, query processing, retrieval, context optimization, synthesis, quality assurance, post-processing
• Chunking strategies: recursive, semantic, token, fixed_size, sliding_window
• Retrieval: Vector search with top_k / similarity_threshold controls
• Reranking: FlashRank cross-encoder reranker
• Synthesis: Direct, extractive, abstractive, and hybrid synthesizers
• HyDE support: Hypothetical Document Embeddings for query expansion
• Citations: Inline, footnote, or numbered citation styles
• Quality assurance: Faithfulness check and hallucination detection

Testing

async with Application.boot(modules=[RAGModule.stub()]) as app:
    # your test code
    ...

Key Source Files

Multi-Tenancy

lexigram-ai-rag supports per-tenant pipeline isolation. When tenancy is enabled, the provider registers a TenantScopedRAGPipeline — a caching wrapper that builds a dedicated RAGPipelineProtocol per tenant, with a tenant-resolved collection_name.

Note: Enabling tenancy requires the app-wide Context binding, which ships with the core bootstrap module (CoreModule / StandardModule from lexigram.app). A bare Application.boot(modules=[RAGModule...]) without it fails with UnresolvableDependencyError: Context.

Configuration

from lexigram.ai.rag import RAGModule
from lexigram.ai.rag.config import RAGConfig, RAGTenancyConfig

config = RAGConfig(
    tenancy=RAGTenancyConfig(enabled=True),
    collection_name="my_docs",
)
RAGModule.configure(config)

Per-Tenant Collection Name

Use RAGConfig.with_collection() to create configs scoped to different collection names — handy for dynamic per-tenant pipeline construction:

tenant_config = RAGConfig().with_collection("tenant_a_collection")

Components